基于多特征组合的图像纹理分类

摘要

在对纹理图像进行特征提取的算法中,高斯马尔可夫随机场(GMRF)、局部二值模式(LBP)和灰度共生矩阵(GLCM)这三种算法应用的较为广泛.常见的图像纹理分类做法是取某一种特征提取算法得到各种纹理的特征空间,进而配合分类算法进行分类.然而,这种做法的不足之处在于未能充分利用各种特征之间的关联,且选取某一种特征提取算法建立特征空间不具对比性.对此,提出一种多特征组合的方法,通过比较单个算法特征与组合特征的分类效果探究各算法特征在对纹理图像的分类上是否存在信息互补.实验结果表明单个算法特征在纹理分类上的确存在优势互补,实验中所得最佳组合特征将给定图像纹理的平均分类精度提高到96.9%.%Among texture image feature extraction algorithms, Gaussian Markov Random Field (GMRF), Local Binary Patterns (LBP) and Cray Level Co-occurrence Matrix (CLCM) are the three comparatively widely used ones. The common image texture classification method is to choose one kind of feature extraction algorithm to obtain the feature space of various kinds of textures, then to cooperate with the classification algorithm for classification. However there are also weaknesses in the approach. On the one hand it doesn' t fully use the relations among all kinds of features, on the other hand it lacks comparison in choosing one kind.of feature extraction algorithm to build the feature space. Therefore a feature combination approach is proposed, which, by comparing the classification effects between the single algorithm feature and the combined feature, to explore whether there are mutual complementary information algorithm features on texture image classification. Experimental results demonstrate that the single algorithm feature truly possesses advantage complementation on texture classification while the optimal combination feature obtained from the experiment has improved the designated image feature' s average classification accuracy to 96.9%.